CN116304518A - Heterogeneous graph convolution neural network model construction method and system for information recommendation - Google Patents

Heterogeneous graph convolution neural network model construction method and system for information recommendation Download PDF

Info

Publication number
CN116304518A
CN116304518A CN202310389653.2A CN202310389653A CN116304518A CN 116304518 A CN116304518 A CN 116304518A CN 202310389653 A CN202310389653 A CN 202310389653A CN 116304518 A CN116304518 A CN 116304518A
Authority
CN
China
Prior art keywords
neural network
network model
heterogeneous
information
matrix
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310389653.2A
Other languages
Chinese (zh)
Inventor
王昊
李际超
姚锋
宋彦杰
杨文川
欧俊威
雷天扬
王涛
陈英武
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
National University of Defense Technology
Original Assignee
National University of Defense Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by National University of Defense Technology filed Critical National University of Defense Technology
Priority to CN202310389653.2A priority Critical patent/CN116304518A/en
Publication of CN116304518A publication Critical patent/CN116304518A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computational Mathematics (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Operations Research (AREA)
  • Probability & Statistics with Applications (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention provides a heterogeneous graph convolution neural network model construction method and system for information recommendation, wherein the method comprises the following steps: obtaining resume information of a plurality of early job seekers, and constructing a heterogeneous resume graph based on all resume information; extracting a feature matrix from the heterogeneous resume graph; acquiring association attributes among the resume information, and constructing an initial adjacency matrix based on the association attributes; constructing a plurality of element paths taking the resume information as nodes according to the education information and the skill information in the resume information; aggregating all of the meta paths and the initial adjacency matrix into a target adjacency matrix; and taking the characteristic matrix and the target adjacent matrix as model input, and constructing a heterogeneous graph convolution neural network model by taking the graph convolution neural network as a framework. The heterogeneous graph convolutional neural network model constructed by the invention can save time and energy of early job seekers in the recruitment information screening process.

Description

Heterogeneous graph convolution neural network model construction method and system for information recommendation
Technical Field
The invention belongs to the technical field of Internet, and particularly relates to a heterogeneous graph convolution neural network model construction method and system for information recommendation.
Background
In recent years, with the development of related technologies of artificial intelligence, an online recruitment platform is continuously emerging, which aims to provide services for potential job seekers and personnel units. In view of the job availability and the numerous job seekers, it is difficult for the recruiter to find the appropriate job seeker in time. Typically, the recruiter will be more concerned about the two parts of the early job seeker's resume, namely the educational background and skill level. The educational background reflects the habit of the job seeker to maintain long-term effort, while the skill level reflects the ability of the job seeker to learn new things. The traditional recruitment process mainly depends on subjective experience of recruiters, and quantitative and comprehensive evaluation of resume information of a large number of job seekers is difficult. Therefore, many recruitment platforms construct a neural network model based on a machine learning method, so that person post matching is performed rapidly, and appropriate job seeker resume is pushed to recruiters, so that intelligent human resource management is realized.
However, recruitment is a two-way selection process, and the two-way requirements of both the job seeker and the recruiter should be met, and recommending a proper position to the job seeker is also an extremely critical task of the recruitment platform. However, the research of the machine learning method adopted by the existing recruitment platform is seriously dependent on the manual characteristics and expert knowledge, so that the cost is high, the updating is difficult and the error is easy to occur. The existing recruitment platforms using the machine learning model are all used for recommending recruitment information according to the working experience of the job seeker, and the existing machine learning model cannot be directly applied to the early job seekers such as the candidate graduation because the early job seeker does not have the corresponding working experience. In the face of massive recruitment information, early job seekers had to spend a great deal of time and effort searching and screening to find a relatively suitable job.
Disclosure of Invention
The invention provides a heterogeneous graph convolution neural network model construction method and system for information recommendation, which are used for solving the problem that a machine learning model cannot be applied to early job seekers, so that the early job seekers can find a relatively suitable working position only by spending a great deal of time and effort to screen massive recruitment information.
In a first aspect, the present invention provides a method for constructing a hetero-graph convolutional neural network model for information recommendation, the method comprising the steps of:
obtaining resume information of a plurality of early job seekers, and constructing a heterogeneous resume graph based on all resume information;
extracting a feature matrix from the heterogeneous resume graph;
acquiring association attributes among the resume information, and constructing an initial adjacency matrix based on the association attributes;
constructing a plurality of element paths taking the resume information as nodes according to the education information and the skill information in the resume information;
aggregating all of the meta paths and the initial adjacency matrix into a target adjacency matrix;
and taking the characteristic matrix and the target adjacent matrix as model input, and constructing a heterogeneous graph convolution neural network model by taking the graph convolution neural network as a framework.
Optionally, the heterogeneous graph rolled neural network model further comprises two hidden layers and one output layer.
Optionally, the functional expression of the neural network layer in the heterogram convolutional neural network model is:
f(H (l) ,B)＝RELU(BH (l) W (l) )
wherein: h (0) ＝X，H (l) =z, X is the feature matrix, l is the number of layers of the neural network layer, Z is the level output, W (l) For the weights of the neural network layer, RELU (·) is a nonlinear activation function RELU, and RELU (x) =max (0, x).
Optionally, the output layer is a softmax layer, and the expression of the output layer is:
Figure BDA0004175361070000021
optionally, a loss function in the heterogram convolutional neural network model is defined as a negative log likelihood loss, and an expression of the loss function is:
Figure BDA0004175361070000022
optionally, the method further comprises the steps of:
and setting the self-circulation of the heterogeneous resume graph by adding a preset identity matrix and the target adjacent matrix.
Optionally, the method further comprises the steps of:
normalizing the target adjacency matrix to a symmetric adjacency matrix;
and updating a neural network layer in the heterogeneous graph convolutional neural network model based on the symmetrical adjacency matrix.
Optionally, the calculation formula of the symmetric adjacency matrix is:
Figure BDA0004175361070000023
wherein:
Figure BDA0004175361070000024
b is the target adjacency matrix,>
Figure BDA0004175361070000025
for the symmetrical adjacency matrix->
Figure BDA0004175361070000026
A diagonal node degree matrix for the symmetric adjacency matrix.
Optionally, updating the expression of the neural network layer is as follows:
Figure BDA0004175361070000027
wherein: h (0) X, X is the feature matrix, l is the number of layers of the neural network layer, W (l) And sigma is the output layer of the heterogram convolution neural network model for the weight of the neural network layer.
In a second aspect, the present invention also provides a heterogeneous graph convolutional neural network model building system for information recommendation, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method as described in the first aspect when executing the computer program.
The beneficial effects of the invention are as follows:
the heterogeneous graph convolution neural network model construction method for information recommendation comprises the following steps: obtaining resume information of a plurality of early job seekers, and constructing a heterogeneous resume graph based on all resume information; extracting a feature matrix from the heterogeneous resume graph; acquiring association attributes among a plurality of resume information, and constructing an initial adjacency matrix based on the association attributes; constructing a plurality of element paths taking the resume information as nodes according to the education information and the skill information in the resume information; aggregating all of the meta paths and the initial adjacency matrix into a target adjacency matrix; and taking the characteristic matrix and the target adjacent matrix as model input, and constructing a heterogeneous graph convolution neural network model by taking the graph convolution neural network as a framework. Education and skill related information in the resume of the early job seeker can be mined through the heterogeneous graph convolution neural network model, so that the category of the early job seeker is accurately positioned, and the most appropriate recruitment information can be recommended according to the category of the early job seeker, so that time and energy of the early job seeker in the recruitment information screening process are saved. Meanwhile, the process of screening proper early job seekers from the huge job seekers is accelerated.
Drawings
Fig. 1 is a schematic flow chart of a method for constructing a heterogeneous graph convolutional neural network model for information recommendation in one embodiment of the present application.
Fig. 2 is a schematic diagram of a relationship between a parameter dropout and a model accuracy of a heterogeneous convolution neural network in one embodiment of the present application.
Fig. 3 is a schematic diagram of a relationship between a parameter learning rate and a model accuracy of a heterogeneous convolution neural network according to one embodiment of the present application.
FIG. 4 is a schematic diagram of the accuracy of adding different meta-paths to an adjacency matrix in one embodiment of the present application.
Detailed Description
The invention discloses a heterogeneous graph convolution neural network model construction method for information recommendation.
In one embodiment, referring to fig. 1, the method for constructing a heterogeneous graph convolution neural network model for information recommendation specifically includes the following steps:
s101, resume information of a plurality of early job seekers is obtained, and a heterogeneous resume graph is constructed based on all resume information.
Wherein, the early job seeker refers to the seeker who does not have work experience such as the graduate, but has certain academic, skill and learning ability, can communicate withAnd acquiring resume information of the early job seeker through the recruitment platform. The heterogeneous resume graph belongs to a heterogeneous graph, and the heterogeneous graph is defined as
Figure BDA0004175361070000031
Wherein->
Figure BDA0004175361070000032
Is a node set, epsilon represents an edge set, and the network mode of the heterogeneous graph is expressed as
Figure BDA0004175361070000033
The mapping function of node set to node type set is expressed as +.>
Figure BDA0004175361070000034
And the mapping of the link set to the link type set is denoted +.>
Figure BDA0004175361070000035
For each node V e V, there is associated with a typical node type +.>
Figure BDA0004175361070000036
Furthermore, each link E E is associated with a typical link type>
Figure BDA0004175361070000037
The set of node types and link types are denoted +.>
Figure BDA0004175361070000038
And->
Figure BDA0004175361070000039
Wherein->
Figure BDA00041753610700000310
In the present embodiment, the heterogeneous resume graph is defined as g= (a) G ,L G )。A G Is a combination of vertices of different types, L G Is a collection of various links that are linked,wherein l G ∈L G ，a G ∈A G . Assuming that each node is connected to itself, i.e. for any a G Satisfy (a) G ,a G )∈A G . Defining a feature matrix as
Figure BDA00041753610700000311
Where m is the dimension of the feature vector and n is the number of nodes. Every row->
Figure BDA00041753610700000312
Is the eigenvector of a.
S102, extracting a feature matrix from the heterogeneous resume graph.
Each resume information in the heterogeneous resume graph represents a node, and each node has node characteristics. The node characteristics are set as an N X M dimension characteristic matrix X, wherein N is the node number, and M is the node characteristic number.
S103, acquiring association attributes among resume information, and constructing an initial adjacency matrix based on the association attributes.
In this embodiment, the association attribute is a relationship attribute between users on the recruitment platform, and the association attribute includes other user lists displayed on the resume information page of the target user of the recruitment platform, that is, other users in the other user lists are also seen when the resume information page of the target user is viewed. The resume information of other users in the other user list is similar to the resume information of the target user. The association attribute between two resume information can be regarded as a link between the two resume information, so that the initial adjacency matrix B can be constructed from the association attribute. Simultaneously defining a degree matrix D of the initial adjacent matrix B, wherein the expression of the degree matrix D is D pp ＝∑ p B pq . The degree matrix is a common concept in graph theory and is generally used to describe the degree statistics of a graph. In an undirected graph, the degree of each node refers to the number of nodes immediately adjacent to the node, and the degree matrix is a diagonal matrix, where the elements on each diagonal represent the degree of the corresponding node. Due to the existence of self-loops, pairs of initial adjacency matrices BThe corner line elements are each assigned 1.
In one embodiment, the self-loop of the heterogeneous resume graph is set by adding a preset identity matrix to the target adjacency matrix. Specifically, in the representation method of the adjacency matrix, a weight is provided between each node and each adjacent node to represent the connection strength between the nodes, but if the connection between the node and the adjacent node is lost, the loss can be compensated by adding a weight representing self-circulation. The self-circulation weight can be realized by adding an identity matrix in the adjacent matrix.
The method comprises the specific steps of adding the adjacent matrix and the identity matrix to obtain a weighted adjacent matrix, and then taking the weighted adjacent matrix as the input of a self-circulation-based neural network layer, so that self-circulation connection can be introduced into the network. Since the identity matrix represents a connection between a node and itself, the purpose of setting self-loops is achieved by adding the adjacency matrix to correct the adjacency matrix.
S104, constructing a plurality of element paths taking each resume information as a node according to the education information and the skill information in the resume information.
Wherein the meta-path defines nodes
Figure BDA0004175361070000041
And->
Figure BDA0004175361070000042
Relation set between->
Figure BDA0004175361070000043
Wherein the symbol->
Figure BDA0004175361070000044
Representing the combine operators on the relationship. The meta-path expands the concept of link types in heterogeneous graph networks and carries different semantics depending on their composition.
In the present embodiment, res= { r is used 1 ,r 2 ,...,r n N early seeks are representedResume information for a staff, each individual resume information may be described by a set of attributes, denoted as res= { Edu, sk }, where Edu = { e 1 ,e 2 ,...,e t Education information representing early job seekers, the smaller the t value, the earlier the education experience representing early job seekers, and thus e t Can represent the educational experience of the early job seeker with the closest time distance. Educational information may also be subdivided, specifically denoted Edu = { Deg, unr, maj }, where Deg represents the academic information in the resume information, unr represents the school world ranking information in the resume information, and Maj represents the professional information in the resume information. Sk is skill information of early job seekers, defined as Sk= { s 1 ,s 2 ,...,s j }。
Because the overall resume information is too complex and scattered. Therefore, resume information can be processed as classification data, expressed as job= { T 1 ,T 2 ,...,T o Resume information may be described by a set of attributes, denoted res= { Deg, unr, maj, sk }. The relationship attribute between two resume information is expressed as
Figure BDA0004175361070000051
Wherein r is * Attribute information representing resume information.
In this embodiment, different meta paths capture semantic relationships between resume information from different angles, for example, meta path "RDR" indicates that early job seekers corresponding to two resume information have the same level of learning. The meta path "RSR" indicates that two resume information have similar skill information. The meta path "RSR" indicates that two resume information have the same school world ranking information. The meta path "RMR" indicates that two resume information have the same professional information.
S105, aggregating all element paths and the initial adjacency matrix into a target adjacency matrix.
Wherein after constructing the initial adjacency matrix and the different meta paths between resume information, they can be aggregated into a new target adjacency matrix. In this embodiment, the following steps may be specifically adopted:
constructing a meta-path matrix: the graph is subjected to meta-path sampling to obtain a plurality of meta-paths (meta-paths), and then a meta-path matrix is constructed, wherein each column represents one meta-path, and each row represents one node. In the meta-path matrix, each element represents whether a certain meta-path passes through the node, and 0/1 coding is adopted.
Obtaining a node context vector: multiplying the element path matrix by the adjacent matrix to obtain a node context vector matrix. In the node context vector matrix, each row represents a node, and each column represents a node context vector corresponding to a corresponding element path.
Meta-path aggregation: for each node, its context vector is aggregated in a certain way into one vector. Common aggregation approaches include adding all context vectors, averaging, employing attention mechanisms, etc.
Obtaining a new adjacency matrix: the vectors obtained by aggregating each node are combined in a certain way to obtain a new node representation vector as one row (or one column) of a new adjacency matrix.
S106, taking the feature matrix and the target adjacent matrix as model input, and constructing a heterogeneous graph convolution neural network model by taking the graph convolution neural network as a framework.
The heterogeneous graph convolutional neural network model is a heterogeneous graph convolutional network model of semi-supervised node classification. The graph roll-up neural network will aggregate the feature information from the first-order neighbors of the node. Thus, the heterograph convolution neural network model may capture neighbor information through a two-layer convolution. The heterograph convolution neural network model runs directly on the graph and derives node embedding vectors from their neighborhood attributes. When stacking multiple graph roll-up neural network layers, information about a larger neighborhood is aggregated.
In one embodiment, the heterogeneous graph convolutional neural network model constructed in step S106 includes an input layer, two hidden layers, and an output layer, where the feature matrix and the target adjacency matrix are used as input layers, and the node embedments of the two hidden layers have the same size as the tag set. In this embodiment, each neural network layer in the heterograph convolution neural network model may be represented as a nonlinear function, where the functional expression of the neural network layer is:
f(H (l) ,B)＝RELU(BH (l) W (l) )
wherein: h (0) ＝X，H (l) The number of layers of the neural network layer is Z, X is the feature matrix, and Z is the image level output, W (l) As weights of the neural network layer, RELU (·) is a nonlinear activation function RELU, and RELU (x) =max (0, x).
In this embodiment, the output layer is a softmax layer, and the expression of the output layer is:
Figure BDA0004175361070000061
after inputting the heterogeneous resume graph into the heterogeneous graph convolution neural network model, respectively obtaining node embedded vectors under different element paths, then aggregating all the node embedded vectors to obtain a final embedded vector, and finally inputting the final embedded vector into a softmax layer to obtain a graph-level output Z, wherein the expression is as follows:
Figure BDA0004175361070000062
wherein X is the feature matrix of the device,
Figure BDA0004175361070000063
and->
Figure BDA0004175361070000064
Are weights of hidden layers with feature maps, σ represents the softmax layer.
In one embodiment, since the heterogram convolutional neural network model constructed in the present invention is used to solve the multi-classification problem, the loss function is defined as a negative log likelihood loss, and the expression of the loss function is:
Figure BDA0004175361070000065
wherein y is i Representing the calculation of only the logarithm C of the probability value corresponding to the true one of the categories. For example, if c=2, then
Figure BDA0004175361070000066
Figure BDA0004175361070000067
The probability distribution is followed and summed to 1.
Since the hetero-graph convolutional neural network model is a model that learns the mapping relationship from input to output through training. In the model training process, errors are calculated through a backward propagation algorithm (backprojection) and distributed to the connection weights among each neuron, and then the weights and the bias terms are updated through a gradient descent method, so that the performance of the model is continuously optimized. And updating the neural network layer refers to updating the connection weight and the bias term among the neurons in the neural network, so that the network can gradually adapt to training data, and further the performance and the accuracy rate of the training data are improved. In the deep learning field, updating the neural network layer is generally an iterative process, and the structure and parameter setting of the neural network need to be continuously adjusted in combination with actual data and task requirements so as to achieve the best effect.
In one embodiment, the neural network layer in the heterogram rolled neural network model may be updated by:
the target adjacency matrix is normalized to a symmetric adjacency matrix.
And updating the neural network layer in the heterogeneous graph convolution neural network model based on the symmetrical adjacency matrix.
In this embodiment, the calculation formula of the symmetric adjacent matrix is:
Figure BDA0004175361070000068
wherein:
Figure BDA0004175361070000069
b is a target adjacency matrix,>
Figure BDA00041753610700000610
is a symmetrical adjacency matrix->
Figure BDA00041753610700000611
Is a diagonal node degree matrix of a symmetrical adjacency matrix.
The expression for updating the neural network layer is as follows:
Figure BDA0004175361070000071
wherein: h (0) X, X is the feature matrix, l is the number of layers of the neural network layer, W (l) And sigma is the output layer of the hetero-graph convolution neural network model.
In one embodiment, the feasibility of the heterogram convolutional neural network model of the present invention is verified based on an actual dataset and 8 different baseline models, namely a Graph Convolution Network (GCN), a graph meaning network (GAT), a Support Vector Machine (SVM), a K-nearest neighbor (KNN), adaboost (AB), a Random Forest (RF), logistic Regression (LR) and a Naive Bayes Classifier (NBC), respectively.
GCN is used for semi-supervised learning of graph structured data, which is a more efficient version of convolutional neural networks. It is skillfully designed to extract features from the graph data so that these features can be used for node classification purposes. GAT employs an attention mechanism to achieve more efficient neighbor aggregation than GCN. GAT does not require complex matrix operations or a priori knowledge of the graph structure. GAT assigns different importance to different nodes in the neighborhood by superimposing the self-attention layers during the convolution process.
The SVM handles classification tasks by identifying a separation boundary, called the classification plane, that maximizes the distance between the boundary and the nearest data point on each side, with the purpose of effectively classifying the data points into separate classes. The KNN theory is mature, the thought is simple, and the method can be used for classification and regression. The idea of KNN is that if most of the K nearest (i.e. nearest neighbor in the feature space) samples near a sample belong to a certain class in the feature space, then that sample also belongs to that class.
AB is a method that involves training multiple simple classifiers using training data and then combining them into one powerful classifier. RF is a classifier that contains multiple decision trees and works well for multiple classification problems. The output category of RF is judged by each decision tree in the forest, and the final predicted category is determined by the most selected category.
LR can address multiple classification tasks and is used to represent the probability of an event occurring. LR is easy to deploy, requires minimal computational power, provides fast results, and uses minimal storage resources in the classification process. NBC is one of the most widely used classification algorithms based on Bayesian theorem. The nature of the bayesian classification algorithm is to calculate conditional probabilities.
To evaluate the classification performance of all models, the following five representative evaluation indexes were selected: accuracy, precision, macro-F1 score, and AUC. Wherein accuracy measures the proportion of correct predictions of the model. It is calculated by correctly predicting the tag over the total number of instances. The specific calculation formula of the Accuracy Accuracy is as follows:
Figure BDA0004175361070000072
precision refers to the proportion of items identified as Positive by the model that are truly Positive. It may be calculated by dividing the number of positive labels accurately predicted by the total number of positive predictions. The higher the accuracy, the lower the false positive rate. The specific calculation formula of the precision Precison is as follows:
Figure BDA0004175361070000073
the Macro-F1 score combines precision and recall under the concept of a harmonic mean and finds the best trade-off between these two quantities. The specific calculation formula is as follows:
Figure BDA0004175361070000081
Figure BDA0004175361070000082
Figure BDA0004175361070000083
wherein TP, TN, FP, FN represents true positive, true negative, false positive and false negative, respectively.
AUC represents probability, representing the ability of the classification algorithm to distinguish between positive and negative samples. A higher AUC value indicates that the algorithm is more efficient in placing the positive samples before the negative ones, indicating better classification performance. The specific calculation formula of AUC is as follows:
Figure BDA0004175361070000084
wherein P, N, M represents the number of positive samples, negative samples and classification categories, respectively.
In the present embodiment, an experiment was performed in Pytorch Geometric 1.11.0, and the learning rate was set to 0.01 and dropout was set to 0.2. The model is then trained for over 900 generations and the decay weight is set to 0.0005. To generate the same model initialization learnable parameters each time, a random number seed is set at 42. In this case, the number of hidden heterogeneous graph convolutional neural network model layers is set to 16 units. For the hetero-graph convolutional neural network model, a two-step meta-path is used, the number of meta-paths being 4.
TABLE 1 heterogeneous graph convolution neural network model and baseline model performance data
Accuracy of Precision of Macro-F1score AUC
Heterogeneous graph convolution neural network model 0.8591 0.9011 0.8218 0.8475
GCN 0.7350 0.4905 0.5173 0.6911
GAT 0.6644 0.6532 0.6189 0.6299
SVM 0.5745 0.5209 0.5401 0.5495
KNN 0.5532 0.1829 0.1879 0.5345
AB 0.5390 0.4683 0.5003 0.5382
RF 0.5887 0.1951 0.2105 0.5553
LR 0.6312 0.3206 0.2714 0.5829
NBC 0.0922 0.0768 0.0974 0.5643
Experimental results referring to table 1, the heterogram convolutional neural network model performed best, significantly better than all baseline models. The heterogeneous graph convolutional neural network model without added element paths performs worse than the one with added element paths. This suggests that adding a meta-path to a heterogeneous resume graph may preserve syntactic and semantic relationships between resume information, which may provide additional information in large external resume data. The heterogeneous graph convolutional neural network model is obviously better than the baseline model, which indicates that the heterogeneous graph convolutional neural network model can fully combine the characteristics of nodes and neighbors thereof to analyze resume information. NBC relies on the assumption that the sample properties are independent, with the worst results achieved on the experimental dataset, meaning that the sample properties are correlated. GAT performs worse than the heterogram convolutional neural network model, indicating that the embedding of the GAT's unsupervised resume is not very discernable in the job classification.
The heterograph convolutional neural network model is an improved GCN model, and as a special form of Laplacian smoothing, the new features of the nodes are calculated as weighted averages of the nodes themselves and their second-order neighbors. Therefore, the heterogeneous graph convolution neural network model combines the characteristics of the nodes and the neighbors thereof in the heterogeneous graph to analyze the nodes, so the heterogeneous graph convolution neural network model has good performance in node classification. And as the element path is added into the heterogeneous graph convolution neural network model as a characteristic, the heterogeneous graph can capture the node-node relationship and the global node-element path relationship. Thus, the heterograph convolution neural network model is more interpretable.
In one embodiment, the robustness of the heterogram convolutional neural network model is further evaluated by precision test experiments. Referring to fig. 2, fig. 2 shows the test accuracy at different parameter settings. As dropout becomes larger, the test accuracy gradually decreases, but there is a short rise in average accuracy at 0.5. This suggests that dropout randomly generates relatively more network structure at 0.5.
Referring to fig. 3, in fig. 3, classification performance of a heterogram convolutional neural network model with different learning rates is described. It can be observed that the test accuracy reaches the optimal solution and is relatively robust at a parameter learning rate of 0.01. Furthermore, too high a parameter learning rate does not improve classification performance and may cause oscillations.
Referring to fig. 4, fig. 4 shows the accuracy of adding different meta-paths to the adjacency matrix. The variable a in the X-axis represents the use of the association attribute alone as an adjacency matrix. Also, the variables AD, AM, AU, AS, AMS, ADMSU in the axis represent the addition of meta-paths RDR, RMR, RUR, RSR, RMR and RSR, RDR, RMR, RUR and RSR, respectively, to the associated attributes. The heterogram convolution neural network model achieves optimal performance when the adjacency matrix adds all four element paths to the associated attributes. When the adjacency matrix is a combination of the association attribute and the meta-path RDR, the model performance is reduced relative to no addition. However, by adding the meta-path RDR as an AMSU to the adjacency matrix, the accuracy of the heterogram convolutional neural network model is significantly improved. The result shows that the early job seeker should not only pay attention to the learning level, but also comprehensively evaluate the education background and skill level of the early job seeker.
Those of ordinary skill in the art will appreciate that: the discussion of any of the embodiments above is merely exemplary and is not intended to imply that the scope of the present application is limited to such examples; the technical features of the above embodiments or in the different embodiments may also be combined under the idea of the present application, the steps may be implemented in any order, and there are many other variations of the different aspects of one or more embodiments in the present application as above, which are not provided in details for the sake of brevity.
One or more embodiments herein are intended to embrace all such alternatives, modifications and variations that fall within the broad scope of the present application. Any omissions, modifications, equivalents, improvements, and the like, which are within the spirit and principles of the one or more embodiments in the present application, are therefore intended to be included within the scope of the present application.

Claims (10)

1. The heterogeneous graph convolution neural network model construction method for information recommendation is characterized by comprising the following steps of:
obtaining resume information of a plurality of early job seekers, and constructing a heterogeneous resume graph based on all resume information;
extracting a feature matrix from the heterogeneous resume graph;
acquiring association attributes among the resume information, and constructing an initial adjacency matrix based on the association attributes;
constructing a plurality of element paths taking the resume information as nodes according to the education information and the skill information in the resume information;
aggregating all of the meta paths and the initial adjacency matrix into a target adjacency matrix;
and taking the characteristic matrix and the target adjacent matrix as model input, and constructing a heterogeneous graph convolution neural network model by taking the graph convolution neural network as a framework.
2. The method for constructing a hetero-graph convolutional neural network model for information recommendation of claim 1, wherein the hetero-graph convolutional neural network model further comprises two hidden layers and one output layer.
3. The method for constructing a heterogeneous graph convolutional neural network model for information recommendation according to claim 2, wherein the functional expression of the neural network layer in the heterogeneous graph convolutional neural network model is:
f(H (l) ,B)＝RELU(BH (l) W (l) )
wherein: h (0) ＝X，H (l) =z, X is the feature matrix, l is the number of layers of the neural network layer, Z is the level output, W (l) For the weights of the neural network layer, RELU (·) is a nonlinear activation function RELU, and RELU (x) =max (0, x).
4. The heterogeneous graph rolling neural network model construction method for information recommendation according to claim 2, wherein the output layer is a softmax layer, and the expression of the output layer is:
Figure FDA0004175361050000011
5. the heterogeneous graph convolutional neural network model construction method for information recommendation of claim 1, wherein a loss function in the heterogeneous graph convolutional neural network model is defined as a negative log likelihood loss, and an expression of the loss function is:
Figure FDA0004175361050000012
6. the heterogeneous-graph convolutional neural network model construction method for information recommendation of claim 1, further comprising the steps of:
and setting the self-circulation of the heterogeneous resume graph by adding a preset identity matrix and the target adjacent matrix.
7. The heterogeneous-graph convolutional neural network model construction method for information recommendation of claim 1, further comprising the steps of:
normalizing the target adjacency matrix to a symmetric adjacency matrix;
and updating a neural network layer in the heterogeneous graph convolutional neural network model based on the symmetrical adjacency matrix.
8. The heterogeneous graph convolution neural network model construction method for information recommendation according to claim 7, wherein a calculation formula of the symmetric adjacency matrix is:
Figure FDA0004175361050000021
wherein:
Figure FDA0004175361050000022
b is the target adjacency matrix,>
Figure FDA0004175361050000023
for the symmetrical adjacency matrix->
Figure FDA0004175361050000024
A diagonal node degree matrix for the symmetric adjacency matrix.
9. The heterogeneous graph rolling neural network model construction method for information recommendation of claim 8, wherein updating the expression of the neural network layer is as follows:
Figure FDA0004175361050000025
wherein: h (0) X, X is the feature matrix, l is the number of layers of the neural network layer, W (l) And sigma is the output layer of the heterogram convolution neural network model for the weight of the neural network layer.
10. A heterogram convolutional neural network model building system for information recommendation, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of claims 1 to 9 when executing the computer program.
CN202310389653.2A 2023-04-11 2023-04-11 Heterogeneous graph convolution neural network model construction method and system for information recommendation Pending CN116304518A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310389653.2A CN116304518A (en) 2023-04-11 2023-04-11 Heterogeneous graph convolution neural network model construction method and system for information recommendation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310389653.2A CN116304518A (en) 2023-04-11 2023-04-11 Heterogeneous graph convolution neural network model construction method and system for information recommendation

Publications (1)

Publication Number Publication Date
CN116304518A true CN116304518A (en) 2023-06-23

Family

ID=86778130

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310389653.2A Pending CN116304518A (en) 2023-04-11 2023-04-11 Heterogeneous graph convolution neural network model construction method and system for information recommendation

Country Status (1)

Country Link
CN (1) CN116304518A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116757652A (en) * 2023-08-17 2023-09-15 北京华品博睿网络技术有限公司 Online recruitment recommendation system and method based on large language model

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116757652A (en) * 2023-08-17 2023-09-15 北京华品博睿网络技术有限公司 Online recruitment recommendation system and method based on large language model
CN116757652B (en) * 2023-08-17 2023-10-20 北京华品博睿网络技术有限公司 Online recruitment recommendation system and method based on large language model

Similar Documents

Publication Publication Date Title
US20220391767A1 (en) System and method for relational time series learning with the aid of a digital computer
Sarker Machine learning: Algorithms, real-world applications and research directions
CN110263227B (en) Group partner discovery method and system based on graph neural network
CN111079639B (en) Method, device, equipment and storage medium for constructing garbage image classification model
Fong et al. Accelerated PSO swarm search feature selection for data stream mining big data
Tama et al. Tree-based classifier ensembles for early detection method of diabetes: an exploratory study
Abd-Alsabour A review on evolutionary feature selection
AlFarraj et al. Optimized feature selection algorithm based on fireflies with gravitational ant colony algorithm for big data predictive analytics
US10956825B1 (en) Distributable event prediction and machine learning recognition system
US11151463B2 (en) Distributable event prediction and machine learning recognition system
CN112529638B (en) Service demand dynamic prediction method and system based on user classification and deep learning
CN112632984A (en) Graph model mobile application classification method based on description text word frequency
Chen et al. Generative dynamic link prediction
CN115688024A (en) Network abnormal user prediction method based on user content characteristics and behavior characteristics
Gupta et al. A new swarm-based efficient data clustering approach using KHM and fuzzy logic
Protopapadakis et al. An investigation on multi-objective optimization of feedforward neural network topology
CN116304518A (en) Heterogeneous graph convolution neural network model construction method and system for information recommendation
CN114036298B (en) Node classification method based on graph convolution neural network and word vector
Chemchem et al. Deep learning and data mining classification through the intelligent agent reasoning
Degirmenci et al. iMCOD: Incremental multi-class outlier detection model in data streams
Farooq Genetic algorithm technique in hybrid intelligent systems for pattern recognition
Ahan et al. Social network analysis using data segmentation and neural networks
CN107402984B (en) A kind of classification method and device based on theme
Wang et al. A pso-based web document classification algorithm
Nirmalraj et al. Sparrow Search Optimization with Ensemble of Machine Learning Model for Customer Retention Prediction and Classification

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination